仔细看下他的测试方式，这个比较是非常不公平的

what do we measure and how?
we use a 16 node cluster running at sics. we plot throughput vs. parallel load.

machine 1 has a server (apache or yaws).
machine 2 requests 20 kbyte pages from machine 1. it does this in tight a loop requesting a new page as soon as it has received a page from the server. from this we derive a throughput figure, which is plotted in the horizontal scale on the graph. a typical value (800) means the throughput is 800 kbytes/sec.
machines 3 to 16 generate load.
each machine starts a large number of parallel sessions.

each session makes a very slow request to fetch a one byte file from machine 1. this is done by sending very slow http get requests (we break up the get requests and send them character at a time, with about ten seconds between each character)

apache的链接处理机制是 开线程或者进程来处理请求 按它的测试方法 你非常慢速的8w请求 导致apache开大量的线程来处理。而能开多少线程取决于操作系统的能力 这还是其次 大量的线程处理活跃的链接导致大量的thread content switch。 apache 挂了不奇怪。 而erlang的线程相大于c语言的一个数据结构 erl_process你开多少取决于你的内存 大量的但是慢速的链接刚好适合poll事件dispatch， 以epoll的能力（俺测试过epoll30w）能够轻松处理。 这个测试与其说测试web服务器的性能 不如说 测试服务器的进程生成能力。

俺的测试是这样的：.
./yaws --conf yaws.conf --erlarg "+k true +p 1024000"     #epoll 最多1024000个进程 内核都已经调优过

yaws.conf 的内容：

auth_log = false
max_num_cached_files = 8000
max_num_cached_bytes = 6000000

大家都用 ab -c 1000 -n 1000000 http://192.168.0.98:8000/bomb.gif 来测
果然发现yaws的性能也是非常一般 大概也就是3k左右.

各位看下 strace 的结果就知道了：

accept(10, {sa_family=af_inet, sin_port=htons(5644), sin_addr=inet_addr("192.168.0.97")}, [16]) = 11
fcntl64(11, f_getfl)                    = 0x2 (flags o_rdwr)
fcntl64(11, f_setfl, o_rdwr|o_nonblock) = 0
getsockopt(10, sol_tcp, tcp_nodelay, [0], [4]) = 0
getsockopt(10, sol_socket, so_keepalive, [0], [4]) = 0
getsockopt(10, sol_socket, so_priority, [0], [4]) = 0
getsockopt(10, sol_ip, ip_tos, [0], [4]) = 0
getsockopt(11, sol_socket, so_priority, [0], [4]) = 0
getsockopt(11, sol_ip, ip_tos, [0], [4]) = 0
setsockopt(11, sol_ip, ip_tos, [0], 4)  = 0
setsockopt(11, sol_socket, so_priority, [0], 4) = 0
getsockopt(11, sol_socket, so_priority, [0], [4]) = 0
getsockopt(11, sol_ip, ip_tos, [0], [4]) = 0
setsockopt(11, sol_socket, so_priority, [0], 4) = 0
getsockopt(11, sol_socket, so_priority, [0], [4]) = 0
getsockopt(11, sol_ip, ip_tos, [0], [4]) = 0
setsockopt(11, sol_socket, so_keepalive, [0], 4) = 0
setsockopt(11, sol_ip, ip_tos, [0], 4)  = 0
setsockopt(11, sol_socket, so_priority, [0], 4) = 0
getsockopt(11, sol_socket, so_priority, [0], [4]) = 0
getsockopt(11, sol_ip, ip_tos, [0], [4]) = 0
setsockopt(11, sol_tcp, tcp_nodelay, [0], 4) = 0
setsockopt(11, sol_socket, so_priority, [0], 4) = 0
recv(11, "get /bomb.gif http/1.0\r\nuser-age"..., 8192, 0) = 100
getpeername(11, {sa_family=af_inet, sin_port=htons(5644), sin_addr=inet_addr("192.168.0.97")}, [16]) = 0
clock_gettime(clock_monotonic, {110242, 326908594}) = 0
stat64("/var/www/html/bomb.gif", {st_mode=s_ifreg|0644, st_size=4096, ...}) = 0
access("/var/www/html/bomb.gif", r_ok)  = 0
access("/var/www/html/bomb.gif", w_ok)  = 0
clock_gettime(clock_monotonic, {110242, 327135982}) = 0
time(null)                              = 1185894828
clock_gettime(clock_monotonic, {110242, 327222643}) = 0
stat64("/etc/localtime", {st_mode=s_ifreg|0644, st_size=405, ...}) = 0
writev(11, [{null, 0}, {"http/1.1 200 ok\r\nconnection: clo"..., 231}, {"\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"...
, 4096}], 3) = 4327
close(11

这里面充斥着大量的无用的昂贵的系统调用 （至少有20个*10us = 200us 的系统调用是无效的）
对文件的access 2 次  连文件的cache都没有  每次 打开文件  读文件 然后写到socket去 。

这个case是小文件(4k)的情况。 看下大文件(40k)的情况
open("/var/www/html/bomb.gif", o_rdonly|o_largefile) = 19
read(19, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 10240) = 10240
writev(16, [{null, 0}, {"\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 10240}], 2) = 10240
read(19, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 10240) = 10240
writev(16, [{null, 0}, {"\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 10240}], 2) = 10240
read(19, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 10240) = 10240
writev(16, [{null, 0}, {"\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 10240}], 2) = 10240
read(19, "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 10240) = 10240
writev(16, [{null, 0}, {"\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0"..., 10240}], 2) = 7240
read(19, "", 10240)                     = 0
close(19)                               = 0
clock_gettime(clock_monotonic, {110574, 856508319}) = 0
epoll_ctl(3, epoll_ctl_del, 11, {0, {u32=11, u64=581990243524149259}}) = 0
epoll_ctl(3, epoll_ctl_del, 12, {0, {u32=12, u64=581990243524149260}}) = 0
epoll_ctl(3, epoll_ctl_add, 16, {epollout, {u32=16, u64=581990243524149264}}) = 0
epoll_wait(3, {}, 256, 0)               = 0
clock_gettime(clock_monotonic, {110574, 856677411}) = 0
clock_gettime(clock_monotonic, {110574, 856729274}) = 0